Visualizing high-dimensional predictive model quality.pdfVIP

Visualizing High-Dimensional Predictive Model Quality Penny Rheingans Marie desJardins University of Maryland, Baltimore County SRI International Department of Computer Science and Electrical Engineering Arti?cial Intelligence Center rheingan@ marie@ Abstract Using inductive learning techniques to construct classi?cation mod- els from large, high-dimensional data sets is a useful way to make In classi?cation problems, one of the variables is a distinguished predictions in complex domains. However, these models can be dif- class variable; we refer to the other variables as input variables. ?cult for users to understand. We have developed a set of visualiza- (The class variable can be thought of as the dependent variable; tion methods that help users to understand and analyze the behavior the input variables as the independent variables.) The data space of learned models, including techniques for high-dimensional data is the n-dimensional space de?ned by the n input variables. In a space projection, display of probabilistic predictions, variable/class classi?cation task, the goal is to derive the class probabilities, i.e., correlation, and instance mapping. We show the results of apply- the marginal probabilities that an instance belongs to each class, ing these techniques to models constructed from a benchmark data given values for some (or all) of the input variables. set of census data, and draw conclusions about the utility of these The problem of accurately predicting class membership from methods for model understanding. available information is a key challenge of knowledge discovery. A wide variety of methods have been developed by machine learn-